Multivariate data mapping based on dendritic lattice associative memories

Abstract

We describe a dendritic lattice hetero-associative memory (DLHAM) that performs multivariate numerical data mapping with respect to a set of prototype data vectors selected by diverse objective or subjective criteria. The memory is a feedforward four-layer dendritic neural network based on lattice algebra operations that computes the nearest match between input and prototype data vectors. Our approach shows the inherent capability of n-dimensional vector association to realize coarse or fine data mapping that is computationally simple. Specifically, we apply the DLHAM in a two stage algorithm to the quantization and transfer of Red-Green-Blue (RGB) color coded images. Input color pixels are first quantized and then the resulting representative colors are mapped to another set of palette colors by hetero-association. Examples and quantization error are included to show the DLHAM performance.